概述

在RocksDB中MANIFEST保存了存储引擎的内部的一些状态元数据，简单来说当系统异常重启，或者程序异常被退出之后，RocksDB需要有一种机制能够恢复到一个一致性的状态， 而这个一致性的状态就是靠MANIFEST来保证的.

MANIFEST在RocksDB中是一个单独的文件，而这个文件所保存的数据基本是来自于VersionEdit这个结构.

MANIFEST包含了两个文件，一个log文件一个包含最新MANIFEST文件名的文件，Manifest的log文件名是这样 MANIFEST-(seqnumber),这个seq会一直增长.只有当 超过了指定的大小之后，MANIFEST会刷新一个新的文件,当新的文件刷新到磁盘(并且文件名更新)之后，老的文件会被删除掉.这里可以认为每一次MANIFEST的更新都代表一次snapshot.

下面就是MANIFEST的基本文件组成:

  1. MANIFEST = { CURRENT, MANIFEST-<seq-no>* } 
  2. CURRENT = File pointer to the latest manifest log
  3. MANIFEST-<seq no> = Contains snapshot of RocksDB state and subsequent modifications

在RocksDB中任意时间存储引擎的状态都会保存为一个Version(也就是SST的集合)，而每次对Version的修改都是一个VersionEdit,而最终这些VersionEdit就是 组成manifest-log文件的内容.

下面就是MANIFEST的log文件的基本构成:

  1. version-edit      = Any RocksDB state change
  2. version           = { version-edit* }
  3. manifest-log-file = { version, version-edit* }
  4.                   = { version-edit* }

实现

整个MANIFEST涉及到三个数据结构分别是VersionEdit/Version/VersionSet,其中前两个上面已经有介绍，而最后一个VersionSet顾名思义表示一堆Version的集合，其实就是 记录了各个版本的信息用来管理整个Version.

  1. class VersionSet {
  2.  public:
  3.   VersionSet(const std::string& dbname, const ImmutableDBOptions* db_options,
  4.              const EnvOptions& env_options, Cache* table_cache,
  5.              WriteBufferManager* write_buffer_manager,
  6.              WriteController* write_controller);
  7.   ~VersionSet();
  8. .......................
  9.  private:
  10.   struct ManifestWriter;
  12.   friend class Version;
  13. .................................
  14.   // Opened lazily
  15.   unique_ptr<log::Writer> descriptor_log_;
  16.   // generates a increasing version number for every new version
  17.   uint64_t current_version_number_;
  19.   // Queue of writers to the manifest file
  20.   std::deque<ManifestWriter*> manifest_writers_;
  21. ..........................................

这里最关键的两个数据结构是descriptor_log_和manifest_writers_,前一个表示了当前manifest-log文件，后一个表示需要写入到manifest-log文件中的内容.

下面就是ManifestWriter的结构，可以看到其中包含了一个VersionEdit的数组，这个数组就是即将要写入到manifest-log文件中的内容.

  1. // this is used to batch writes to the manifest file
  2. struct VersionSet::ManifestWriter {
  3.   Status status;
  4.   bool done;
  5.   InstrumentedCondVar cv;
  6.   ColumnFamilyData* cfd;
  7.   const autovector<VersionEdit*>& edit_list;
  9.   explicit ManifestWriter(InstrumentedMutex* mu, ColumnFamilyData* _cfd,
  10.                           const autovector<VersionEdit*>& e)
  11.       : done(false), cv(mu), cfd(_cfd), edit_list(e) {}
  12. };

然后我们来看RocksDB如何来创建以及写入文件,下面所有的代码都是包含在VersionSet::LogAndApply这个函数中.

首先在每次LogAndApply的时候都会创建一个新的ManifesWriter加入到manifest_writers_队列中.这里只有当之前保存在队列中 的所有Writer都写入完毕之后才会加入到队列，否则就会等待.

  1.   // queue our request
  2.   ManifestWriter w(mu, column_family_data, edit_list);
  3.   manifest_writers_.push_back(&w);
  4.   while (!w.done && &w != manifest_writers_.front()) {
  5.     w.cv.Wait();
  6.   }
  7.   if (w.done) {
  8.     return w.status;
  9.   }

接下来就是保存对应的数据到batch_edits中(manifest_writers_).

  1. autovector<VersionEdit*> batch_edits;
  2. ....................................
  3.  if (w.edit_list.front()->IsColumnFamilyManipulation()) {
  4.     // no group commits for column family add or drop
  5.     LogAndApplyCFHelper(w.edit_list.front());
  6.     batch_edits.push_back(w.edit_list.front());
  7.   } else {
  8.     v = new Version(column_family_data, this, current_version_number_++);
  9. ........................................................
  10.     for (const auto& writer : manifest_writers_) {
  11.       if (writer->edit_list.front()->IsColumnFamilyManipulation() ||
  12.           writer->cfd->GetID() != column_family_data->GetID()) {
  13.         break;
  14.       }
  15.       last_writer = writer;
  16.       for (const auto& edit : writer->edit_list) {
  17. ...........................................
  18.         batch_edits.push_back(edit);
  19.       }
  20.     }
  21.     builder->SaveTo(v->storage_info());
  22.   }

然后就是创建新的manifest-log文件的逻辑.这里可以看到要么是第一次进入，要么文件大小大于option对应的值才会创建新的文件

  1. if (!descriptor_log_ ||
  2.       manifest_file_size_ > db_options_->max_manifest_file_size) {
  3.     pending_manifest_file_number_ = NewFileNumber();
  4.     batch_edits.back()->SetNextFile(next_file_number_.load());
  5.     new_descriptor_log = true;
  6.   } else {
  7.     pending_manifest_file_number_ = manifest_file_number_;
  8.   }
  10.   if (new_descriptor_log) {
  11.     // if we're writing out new snapshot make sure to persist max column family
  12.     if (column_family_set_->GetMaxColumnFamily() > 0) {
  13.       w.edit_list.front()->SetMaxColumnFamily(
  14.           column_family_set_->GetMaxColumnFamily());
  15.     }
  16.   }

如果需要创建新的manifest-log文件，则开始构造对应的文件信息并创建文件.

  1. if (new_descriptor_log) {
  2.       // create manifest file
  3.       ROCKS_LOG_INFO(db_options_->info_log, "Creating manifest %" PRIu64 "\n",
  4.                      pending_manifest_file_number_);
  5.       unique_ptr<WritableFile> descriptor_file;
  6.       EnvOptions opt_env_opts = env_->OptimizeForManifestWrite(env_options_);
  7.       s = NewWritableFile(
  8.           env_, DescriptorFileName(dbname_, pending_manifest_file_number_),
  9.           &descriptor_file, opt_env_opts);
  10.       if (s.ok()) {
  11.         descriptor_file->SetPreallocationBlockSize(
  12.             db_options_->manifest_preallocation_size);
  14.         unique_ptr<WritableFileWriter> file_writer(
  15.             new WritableFileWriter(std::move(descriptor_file), opt_env_opts));
  16.         descriptor_log_.reset(
  17.             new log::Writer(std::move(file_writer), 0, false));
  18.         s = WriteSnapshot(descriptor_log_.get());
  19.       }
  20.     }

开始写入对应的VersionEdit的record到文件(最后我们会来看这个record的构成),这里看到写入完成后会调用Sync来刷新内容到磁盘,等这些操作都做完之后，则会更新Current文件也就是更新最新的manifest-log文件名到CURRENT文件中.

  1.       for (auto& e : batch_edits) {
  2.         std::string record;
  3.         if (!e->EncodeTo(&record)) {
  4.           s = Status::Corruption(
  5.               "Unable to Encode VersionEdit:" + e->DebugString(true));
  6.           break;
  7.         }
  8.         TEST_KILL_RANDOM("VersionSet::LogAndApply:BeforeAddRecord",
  9.                          rocksdb_kill_odds * REDUCE_ODDS2);
  10.         s = descriptor_log_->AddRecord(record);
  11.         if (!s.ok()) {
  12.           break;
  13.         }
  14.       }
  15.       if (s.ok()) {
  16.         s = SyncManifest(env_, db_options_, descriptor_log_->file());
  17.       }
  18. .............................
  19.     // If we just created a new descriptor file, install it by writing a
  20.     // new CURRENT file that points to it.
  21.     if (s.ok() && new_descriptor_log) {
  22.       s = SetCurrentFile(env_, dbname_, pending_manifest_file_number_,
  23.                          db_directory);
  24.     }

CURRENT文件更新完毕之后，就可以删除老的mainfest文件了.

  1.   // Append the old mainfest file to the obsolete_manifests_ list to be deleted
  2.   // by PurgeObsoleteFiles later.
  3.   if (s.ok() && new_descriptor_log) {
  4.     obsolete_manifests_.emplace_back(
  5.         DescriptorFileName("", manifest_file_number_));
  6.   }

最后则是更新manifest_writers_队列，唤醒之前阻塞的内容.

  1.   // wake up all the waiting writers
  2.   while (true) {
  3.     ManifestWriter* ready = manifest_writers_.front();
  4.     manifest_writers_.pop_front();
  5.     if (ready != &w) {
  6.       ready->status = s;
  7.       ready->done = true;
  8.       ready->cv.Signal();
  9.     }
  10.     if (ready == last_writer) break;
  11.   }
  12.   // Notify new head of write queue
  13.   if (!manifest_writers_.empty()) {
  14.     manifest_writers_.front()->cv.Signal();
  15.   }

文件内容

具体的文件格式可以看RocksDB的wiki,这里我来介绍下对应的源码. 通过上面的分析我们可以看到最终是通过VersionEdit::EncodeTo来序列化数据,而VersionEdit主要包含了比如log_number/last_sequence_这些字段，这里还有一个比较关键的信息被序列化了，那就是FileMetaData,也就是SST文件的元信息.

  1. struct FileMetaData {
  2.   FileDescriptor fd;
  3.   InternalKey smallest;            // Smallest internal key served by table
  4.   InternalKey largest;             // Largest internal key served by table
  5.   SequenceNumber smallest_seqno;   // The smallest seqno in this file
  6.   SequenceNumber largest_seqno;    // The largest seqno in this file
  8. .........................................
  9.   // File size compensated by deletion entry.
  10.   // This is updated in Version::UpdateAccumulatedStats() first time when the
  11.   // file is created or loaded.  After it is updated (!= 0), it is immutable.
  12.   uint64_t compensated_file_size;
  13.   // These values can mutate, but they can only be read or written from
  14.   // single-threaded LogAndApply thread
  15.   uint64_t num_entries;            // the number of entries.
  16.   uint64_t num_deletions;          // the number of deletion entries.
  17.   uint64_t raw_key_size;           // total uncompressed key size.
  18.   uint64_t raw_value_size;         // total uncompressed value size.
  20.   int refs;  // Reference count
  22.   bool being_compacted;        // Is this file undergoing compaction?
  23.   bool init_stats_from_file;   // true if the data-entry stats of this file
  24.                                // has initialized from file.
  26.   bool marked_for_compaction;  // True if client asked us nicely to compact this
  27.                                // file.
  28. };

工具

这里查看MANIFEST文件依旧是使用ldb工具:

  1. pagefault@god ~/tools/polarbase/data/.rocksdb $ ../../bin/ldb manifest_dump --path=./MANIFEST-000001
  2. --------------- Column family "default"  (ID 0) --------------
  3. log number: 13
  4. comparator: <NO COMPARATOR>
  5. --- level 0 --- version# 0 ---
  6.  11:80860[' 
  7. --------------- Column family "__system__"  (ID 1) --------------
  8. log number: 24
  9. comparator: RocksDB_SE_v3.10
  10. --- level 0 --- version# 1 ---
  11.  25:1094[' 
  12. next_file_number 27 last_sequence 190  prev_log_number 0 max_column_family 1

原文：http://mysql.taobao.org/monthly/2018/05/08/